Hello world, it’s Friday, October 30th.
Nvidia is reportedly in talks to guarantee debt behind Nscale’s $45 billion compute contract with Anthropic. Elsewhere, Samsung posted record third-quarter revenue of 195.3 trillion won on AI memory while its device division lost money again, Intel fell 23% after hours on a weak outlook and supply constraints, and TSMC is close to buying AUO display fabs for CoWoS.
Let’s get into it. — Austin & Vik
Be sure to check out the Semi Doped podcast on YouTube or your favorite podcast player!
Presented by Crusoe
Crusoe builds the full stack for AI, from power and data centers to the cloud. It can also help you tune a small model on your own data for just a few bucks.
For repeatable work like reports, an open model trained on your terminology and formats can match a frontier model, and it’s cheaper and faster to run.
Crusoe’s Serverless Fine-Tuning handles the training: upload your dataset, pick an open model like Llama, DeepSeek, or GLM, and pay per million tokens used, not per GPU hour.
About $2 buys roughly 5M training tokens on smaller models.
Deploy on Crusoe Cloud in one click or download the weights and deploy on your own.
Nvidia Weighs Credit Guarantees to Back Nscale’s Anthropic Compute Deal
Nvidia is in talks to provide credit guarantees on debt tied to Nscale’s contract to supply $45 billion worth of compute to Anthropic, according to The Information and Bloomberg, with the project centered on a data center in West Virginia. Nscale, a neocloud operator, is separately pursuing a $35 billion IPO, according to the Financial Times, making the creditworthiness of this Anthropic contract central to its public-market case. No final terms have been disclosed, and talks are ongoing.
Vik: it's not a good look when NVIDIA needs to backstop purchases to NeoCloud. If NeoCloud (like NScale) isn't getting enough customers, what benefit is there in NVIDIA stepping in, purchasing compute, and renting it back? It's just more bandaid measures.
Austin: I disagree, I think broadly neoclouds are helpful and necessary but access to capital is hard, and like a parent cosigning a mortgage with their first-time homebuying child, Nvidia can help neos get past bank headaches. So long as Nvidia feels the child (neocloud) is capable of making their loan payments, it’s all sensible to me.
Samsung’s Record Profit Runs on Memory as Devices Bleed
Samsung reported third-quarter revenue of 195.3 trillion won and operating profit of 107.4 trillion won, roughly $146 billion and $80 billion respectively, driven almost entirely by surging demand for AI memory. The device division posted another loss, weighed down by higher component costs, leaving memory as the singular engine of a result that looks more lopsided the closer you read it. HBM and advanced DRAM carried the quarter while the chip division itself missed some analyst estimates, according to Bloomberg. Samsung’s management separately warned that the AI chip shortage is worsening, a signal that pricing power in high-bandwidth memory is unlikely to ease soon, though the company kept investors waiting on shareholder return commitments.
Intel Plunges 23% as PC Weakness and Supply Snags Crush Outlook
Intel plunged 23% in late trading after its revenue forecast missed analysts’ estimates, dragged down by weak PC processor demand and supply-chain problems squeezing its data center components. CEO Lip-Bu Tan addressed the shortfall directly in an interview with Bloomberg’s Ed Ludlow, offering little reassurance on either front. The PC slump isn’t new, but supply constraints on data center chips add a sharper edge to an already fragile recovery story: that segment was supposed to be the good news. Cash burn is mounting. Intel’s foundry-and-product turnaround now faces pressure from two directions simultaneously, with AMD and Arm-based alternatives waiting on both the client and server sides of the market.
Vik: people need to stop with this “AI PC”thing. It's not real. Right now AI runs on the cloud and anyone who keeps parroting these four letters is going to continue to lose money
Austin: Bummed to continue to hear Intel unable to keep up with server CPU demand, this has been going on for many quarters. And the other side of the products business seemingly has too much supply into a down client market (due to memory prices, which we’ve known for a long time that memory would hurt client). Am I missing something?
TSMC Nears Deal to Buy Two AUO Fabs for CoWoS Expansion
TSMC is close to acquiring two display fabs from AUO for conversion into CoWoS advanced-packaging capacity, adding to the company’s ongoing expansions at its AP7 and AP8 facilities as demand from AI accelerator customers keeps supply tight. The AUO plant acquisitions would give TSMC additional cleanroom floor space to run chip-on-wafer-on-substrate lines without waiting years for greenfield construction. Substrate and interposer suppliers are feeling the pressure too: Kinsus Technology, a key packaging materials vendor, told Digitimes it faces three distinct challenges scaling up for larger AI chip packages, and the company is exploring aluminum nitride ceramics as a substrate alternative to glass for next-generation designs. Qualification timelines for new materials remain a gating factor across the supply chain.
Vik: I'm mildly intrigued by the proposition to use aluminum nitride instead of glass for substrates. Aluminum nitride has very low loss properties and is already widely used in RF substrates for those properties
Murata and Samsung Electro-Mechanics Race to Supply AI Server Capacitors
Samsung Electro-Mechanics has now signed eight long-term MLCC supply agreements with AI server customers, pushing cumulative contract value past 5 trillion won, with the most recent deal alone worth 879.3 billion won, equivalent to roughly 23% of one quarter’s sales. AI customers are already locking up 2027 supply, according to Digitimes, turning what might look like a procurement cycle into something closer to a land-grab. Murata Manufacturing, Samsung’s chief rival in the segment, is committing ¥1 trillion ($6.4 billion) over five years to more than double its MLCC output for AI data center operators. Both companies are expanding from opposite ends of the same bottleneck: Samsung deepening long-term agreements with named hyperscalers, Murata building the physical capacity to honor commitments at that scale.
Hanmi Semiconductor Hits 50% Operating Margin on HBM TC Bonder Surge
Hanmi Semiconductor reported third-quarter operating profit of 90.8 billion won, a 33.9% year-on-year gain, as surging demand for its thermocompression bonding equipment lifted the company to a 50% operating margin. The margin figure confirms TC bonders as one of the most lucrative chokepoints in the HBM supply chain, with Hanmi capturing pricing power that most equipment suppliers don’t get near. The sequential picture is less tidy: The Elec notes profit fell roughly 30% quarter-on-quarter, though year-on-year momentum remains intact. On the partnership front, SK Hynix formally awarded Hanmi a plaque recognizing their decade-long HBM production alliance, a rare public acknowledgment of a supplier relationship that has deepened alongside HBM4 ramp preparations. Hanmi is separately expanding its packaging business beyond HBM to broaden its revenue base.
Vik: TC bonders stay lucrative only as long as HBM stays on thermocompression and not hybrid bonding - and near term I’m with the crowd that high hybrid bonded stacks cost too much.
Sector Watch
Packaging & Substrates
Ibiden plans a 1.28 trillion won additional investment in its Gama substrate plant to expand CPU and AI package substrate capacity ahead of anticipated 2027 shortages. (www.thelec.net)
Samsung Electro-Mechanics will manufacture ABF substrates in Vietnam to supply AI server and hyperscale data-center customers, expanding capacity outside Korea. (TNGlobal)
Ibiden competitor Unimicron warns PCB equipment lead times have stretched to 30 months with key material shortages likely in 2027, signaling a supply crunch ahead. (XenoSpectrum)
Compute
Foxconn and Nvidia target an eight-week AI factory deployment cycle for Vera Rubin NVL72 racks while Foxconn confirms a new partnership with OpenAI and dismisses AI bubble concerns. (digitimes)
Huawei deploys the first Ascend 950DT 1,024-NPU SuperPoD cluster in Wuhan, completing the build-out in one month as domestic AI compute scales. (Pandaily)
Monolithic Power Systems beats Q3 EPS by $0.48 and appoints a new CFO and Chief Accounting Officer as AI power-management demand accelerates. (StreetInsider)
Memory
Sandisk posts strong Q3 profit and upbeat outlook on growing AI storage demand, marking a sustained NAND recovery after prior inventory correction. (Bloomberg.com)
Seagate books nearline hard-drive production capacity through 2028 as hyperscaler HAMR demand outpaces available supply. (Technetbook)
Micron CEO Sanjay Mehrotra is named recipient of the Global Semiconductor Alliance’s highest honor, the Dr. Morris Chang Exemplary Leadership Award. (Semiconductor Digest)
Foundry
Infineon receives EU approval of €319 million in state aid for its Villach 300mm GaN factory, Europe’s first high-volume GaN wafer line, locking in the €1.1 billion project. (Infineon)
Tokyo Electron raises its fiscal 2026 profit outlook, citing robust leading-edge equipment demand as AI wafer starts continue to ramp. (Bloomberg.com)
Advantest raises its fiscal 2026 revenue and profit outlook as HBM and advanced-logic test demand exceeds prior guidance. (Advantest)
Power
Delta Electronics approves a new Penang manufacturing base to expand global power-supply capacity serving AI data-center customers. (digitimes)
Delta Electronics Thailand reports Q3 sales surging 42.4% year-on-year on AI infrastructure demand, with the full-year outlook pointing higher. (Thailand Business News)
Gaon Cable wins a 120 billion won power-distribution cable order from a US tech giant for an AI data center, the largest cable deal in the LS subsidiary’s history. (The Elec)
Optics & Networking
Corning reports surging profit driven by AI data-center fiber demand, with the CEO maintaining that optical scale-up will accelerate into 2028. (WSJ)
Innolight and Chinese AI optical peers post upbeat Q3 earnings but shares slump on broader fears of slowing AI infrastructure spending. (South China Morning Post)
Lumen launches three new initiatives — Emerald AI, an Ohio State University collaboration, and a Fortinet partnership — to redefine networking for the AI and quantum era. (Light Reading)
Edge & Embedded
Renesas launches the RL78/G21 ultra-low-power MCU targeting smart appliance control, while also reporting Q3 financial results and updated consolidated forecasts. (Business Wire)
STMicroelectronics weighs developing higher-performance MCUs and MPUs for humanoid robots, signaling a push into edge AI robotics silicon. (EE Times)
Wolfspeed signs a silicon carbide supply agreement with BorgWarner to back next-generation EV inverters, extending SiC into automotive power trains. (Wolfspeed)
Policy & Trade
Cohere initially closes a Series E targeting C$7 billion at over $20 billion valuation, backed by Canada’s government and Nvidia, to fund sovereign AI model deployments globally. (Bloomberg.com)
Bloomberg reports Indian AI firm seeks $3 billion in financing to procure Nvidia chips, the latest in a wave of sovereign AI compute funding pushes across Asia. (Bloomberg.com)
Finnish police launch a criminal probe into land clearance practices surrounding Google’s data center, marking the first criminal investigation of a hyperscale site permitting process in Europe. (Reuters)
Tell us how to make Semi Doped better! Give us your feedback! ❤️